Quick summary
At the Apsara Conference on September 22, 2026, Alibaba officially announced that Qwen 4 "is currently in training." As of October 12 there is still no release: no date, price, downloadable weights or API identifier. Yet the internet is full of names, dates and "leaks." Here is what's confirmed and what's rumor.
What's official (Alibaba Cloud press release, Sep 22)
- Qwen 4 is in training.
- The future Qwen 4.5 and Qwen 5 series are projected to scale to 5–10 trillion parameters. That's a long-term roadmap, not a Qwen 4 spec.
- Alibaba reported progress in recursive self-improvement (RSI): over a month of fully automated runs, Qwen3.8-Max completed 33 cycles and raised its Artificial Analysis score from 40 to 45.
- In a chip-design experiment, the model made over 10,000 EDA tool calls in 60 hours and cut a bus module's area by 42% with no performance loss, per the company.
- Alibaba also announced its own AI chips and a goal of operating 20 GW of data center capacity by 2032.
Likely, but not in the press release
Attendees of a talk by Qwen model lead Liu Dayiheng report a "Qwen4 Series Coming Soon" slide with four tiers: Qwen4-Max, Qwen4-Plus, Qwen4-Flash and Qwen4-27B. The 27B would be the open-weight tier for local use. These names come from attendees, not the written release, and no specs (context, license, price) have been published.
In August, Alibaba published Qwen3.8-Flash-Next, presented as an experimental preview of the new architecture (its config uses the qwen4 type, per daily.dev). That's real, but it is not Qwen 4.
Rumors (unconfirmed)
- "Out on October 22": the date appears on an aggregator that admits it is inferred, not an official plan.
- Joint open release of 27B and Max: circulating on X, but we found no official post promising it.
- "Qwen4" files on Hugging Face or download sites: as of Oct 11 there is no Qwen4 repo in the official org. Files with that name are not from Alibaba and may be scams or malware.
What to do meanwhile
- Stay on Qwen 3.8, the current generation.
- Keep your stack model-agnostic with OpenAI-compatible layers so you can switch with one config line.
- Plan ~27B-class hardware for local use; test today with Qwen3.8-27B.
- Download only from official sources (Qwen org on Hugging Face/ModelScope).
- Evaluate on your own data at launch; launch benchmarks don't replace tests on your use case.
Sources
- Alibaba Cloud Apsara press release (Sep 22, 2026): https://www.alibabacloud.com/en/press-room/alibaba-unveils-roadmap-on-full-stack-ai-strategy
- Versely: https://www.versely.studio/blog/qwen-4-announced-at-apsara-2026
- MuseAICode (status as of Oct 11): https://museaicode.bid/models/qwen4/
- daily.dev, speculations: https://daily.dev/posts/speculations-about-qwen4-elqd6an3x
Want to choose and integrate the right AI model for your company? KGA IT offers a free consultation: kga-it.com.